浏览器与自动化
eval-cost
Use when the user asks to make an AI app cheaper or faster, reduce tokens, latency, model/tool/retrieval/rerank/self-check/retry/evaluator cost, or compare cost before/after.
Galileo-Agent-Labs · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add Galileo-Agent-Labs/eval-engineer --skill eval-cost技能说明
name: eval-cost description: Use when the user asks to make an AI app cheaper or faster, reduce tokens, latency, model/tool/retrieval/rerank/self-check/retry/evaluator cost, or compare cost before/after.
Eval Cost
Use this skill for tokenomics RCA. Cost changes are accepted only when Galileo quality metrics do not regress.
Conditional References
- Load
skills/eval-engineer/references/tokenomics-rca.mdwhen choosing the tokenomics workflow or diagnosing why cost moved. - Run
skills/eval-engineer/scripts/compare_tokenomics_packets.pywhen both baseline and verification packets exist. - Load
skills/eval-engineer/assets/cost-diagnosis-template.md,skills/eval-engineer/assets/tokenomics-fix-plan-template.md, andskills/eval-engineer/assets/quality-preserving-verification-template.mdonly when writing those artifacts.
Do
- Compare cost, latency, tokens, retrieved context, tool calls, retries, rerank/self-check spans, model spans, and evaluator cost.
- Run
compare_tokenomics_packets.pywithout explicit quality metrics first when packets use custom quality names, then inspect the inferredQuality metrics comparedlist before accepting the decision. - Treat behavior counters such as handoff count, tool count, step count, retry count, and self-check count as efficiency or workflow evidence, not quality gates by default. Promote one to quality only when the metric profile states the desired direction for that route or segment.
- Protect named quality metrics and segment gates.
- Reject cheaper candidates when aggregate quality holds but any required segment gate regresses.
- Treat lower traffic volume as inconclusive unless per-trace efficiency also improves.
Gotchas
- Cost, latency, wall time, duration, token count, span count, and tool count are efficiency evidence, not quality gates by default.
- Quality metrics are not always higher-is-better. Error, toxicity, hallucination, policy-violation, and tool-error rates regress when they rise.
- Aggregate quality can hide route, risk-profile, or customer-segment regressions.
- For RAG pruning, include hard multi-source or multi-hop cases before keeping a top-k reduction.
- For agentic workflows, compare agent steps, planner spans, rerank passes, self-check spans, and tool calls so the cheaper loop is explainable.
Validation Loop
Before keeping a candidate:
- Compare baseline and verification packets with
compare_tokenomics_packets.py. - Inspect inferred quality metrics before accepting the decision.
- Check segment metrics when available.
- Confirm metric gaps are recorded as inconclusive or follow-up work.
- Reject or revise the candidate if quality evidence is missing or regressed.
Output
State keep/reject/inconclusive first. Explain why the cost moved, which quality metrics do not regress, and which latency/tool-count tradeoffs remain.